August 5's AI news was a day when a technical debate over inference infrastructure and regulatory moves stood out side by side. Cursor, known for its coding assistant, open-sourced a kernel that dramatically speeds up training of large-scale models — even as engineers' voices surfaced claiming that hand-crafted giant kernels (Megakernels) are over. In generative AI, video generation with native audio took a big step toward reality, and on the security front, a large-scale supply chain attack on npm was observed. Here's a roundup of today's major developments.
Cursor Releases "Mixture of Kittens" — and the Same Day, Claims That "the Megakernel Is Over"
Cursor, the coding agent company, has open-sourced "Mixture of Kittens" (MoK), a kernel that accelerates training of MoE (Mixture of Experts) models on NVIDIA's NVL72 systems. By fusing MoE-specific communication and computation into a single kernel, it is reported to be up to 2.37x faster than strong existing open implementations, with a 41% improvement in tokens processed per second. For large-scale services, these are numbers that could translate into cost savings on the order of hundreds of millions of dollars. Development is reportedly led by a former collaborator of Ben Spector, known for the acclaimed kernel "ThunderKittens".
Meanwhile, in inference engineering discussions published around the same time, a strong view emerged that "the Megakernel is dead." The argument: giant hand-fused kernels are hard to write, and in production, the traditional approach — optimizing each step individually and running them in parallel — is often faster after all. On top of that, NVIDIA's next-generation "Rubin" architecture strengthens mechanisms that fill the gaps between kernels (dependency triggers), which suggests the hardware itself is working to solve the very problem that motivated fusion in the first place.
The two developments look contradictory at first glance, but the story holds together once you look at how we got here. What's "over" is mainly giant kernels for inference, while Cursor's MoK targets training. The reading at this point: as hardware advances erode the benefits of fusion on the inference side, training still offers real value in mowing down the communication overhead of large-scale MoE. As the phrase "the Megakernel is dead, and it's back" suggests, the two appear to be fighting on different battlefields.
US Weighs Banning Chinese Data Center Components; Adtek Moves Toward a Major Pre-IPO Raise
On the geopolitical front, it was reported that the United States is considering banning imports of certain Chinese-made data center components, aiming to protect AI infrastructure. Zhongji Innolight, one of the world's leading suppliers of optical transceivers, at one point fell 14% on the Shenzhen market, and peer Eoptolink dropped 11%, as related stocks sold off sharply. The company is reportedly likely to be included in the restrictions. Regulations reaching components that underpin the very foundation of AI data centers seem likely to affect procurement strategies going forward.
Fundraising activity is heating up on the Chinese side as well. Shenzhen-based Adtek Technology is reportedly planning a private raise of up to 2 billion yuan (roughly JPY 29.6 billion) ahead of its Hong Kong IPO. Singapore's state-owned fund Temasek and the private equity arm of Morgan Stanley are said to be among the prospective participants. With regulatory pressure mounting, the sight of Chinese technology companies securing funding before going public is something to keep an eye on.
FLUX 3 Video Goes Native Audio; MiniMax H3 Runs Locally Everywhere
Black Forest Labs, known for image generation, announced "FLUX 3 Video," a video generation model. Its standout feature is native audio handling, and it packs in multilingual dialogue, video generation from text and images, continuation of generation, and a low-cost draft mode. Some are also drawing attention to its ability to predict the next action from footage — a sign that it is stepping beyond mere "can make videos" models toward multimodal generation that includes dialogue and world understanding. Open weights and image-oriented derivative models are slated for release in due course.
Among open video models, MiniMax's recently released "H3" is also spreading fast. Reports of it running on gaming GPUs and MacBooks keep coming, and one engineer documented getting it running locally on an M5 Pro Mac after a roughly 115GB download. In the community, the classic "Will Smith eating spaghetti" test clip became a talking point, with reactions like "well ahead of previous models" (though these are qualitative impressions, with limited details on the settings). Demos using full-precision weights also point to expressive audio and physical consistency — such as sway that changes depending on an object's weight. The speed at which new video models become runnable locally within days is a testament to how mature the ecosystem has become.
Goodfire Releases the Interpretability Platform "Silico"
The field of interpretability — how to make sense of a model's internals — also saw new research infrastructure arrive. Startup Goodfire has released "Silico," an interpretability and training platform for large-scale models. Interpretability work has often lived in notebooks and one-off scripts, which made reproduction and sharing difficult; Silico aims to bring that work together in a shared research environment.
Right after the release, many researchers reported concrete use cases: intervening on Llama and Qwen activations as concept vectors, analyses that reduce unnecessary attention in robotics models, evaluation of ligand binding poses in biochemistry, patch-level recognition of organs and cysts in medical imaging, and reward design to counter guardrail erosion — a wide-ranging picture suggesting that interpretability tools are shifting from mere analysis instruments to a foundation for research and alignment work.
Massive Supply Chain Attack on npm Hits 868 Packages
We also observed an event that sounds the alarm on security. A large-scale supply chain attack was confirmed on npm, with 868 packages affected and more than 2 billion monthly downloads, according to reports. The entry point was compromised maintainer accounts; through preinstall hooks that run at install time, the attack collected credentials for npm, GitHub, AWS, Kubernetes, Vault, and others, and propagated from maintainer to maintainer.
For teams building and operating AI agents and the plugins around them, this kind of event is not someone else's problem. Agents that autonomously call tools and communicate with external systems have the property that failures in dependencies or credentials spread directly into major damage. Separately, technical details of an incident at Hugging Face are reportedly due to be published going forward — which could well become an opportunity to revisit the assumptions behind pulling in third-party code and models.
The "Routing" Trend Reshaping Agent Cost Structures
The "cost versus quality balance" of running agents in production is also taking clear product shape. Not Diamond announced "Not Diamond Code," a router for long-running coding agents that selects the optimal model and reasoning effort at each step, claiming 20–65% cost savings without sacrificing quality. Similarly, the new configuration of "Devin," the autonomous software engineering agent, is reported to have achieved a 4% performance gain and a 27% cost reduction on the "FrontierCode 1.1" benchmark.
In the background, there's also analysis showing that cost per outcome varies 5–30x depending on the choice of harness (the agent's scaffolding). Simple instructions like "just think harder" often do nothing but increase reasoning tokens without contributing to correctness, and design on the scaffolding side — such as comparing multiple approaches — is increasingly pointed to as the biggest lever for efficiency. Alongside the evolution of the models themselves, it seems we've entered a phase where "how you deploy and reuse them" is becoming an axis of competition.
Closing Thoughts
How to speed up inference and training infrastructure, how regulation and geopolitics shake the supply chain, how far multimodal generation gains realism, and how to run agents safely and efficiently. Every one of these topics seems to sit on common ground: "a balance of efficiency and reliability that performance alone can't capture." As new models and tools keep arriving, how to bake cost and safety into design looks set to remain an ongoing question.